Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Parameter Symmetry Potentially Unifies Deep Learning Theory

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Parameter symmetry could unify deep learning theory.

desk verdict A plausible, clearly written position paper that is worth refereeing, but its empirical support rests on thresholded correlations rather than causal tests, and its key theoretical props are mostly the authors' own informal or in-preparation results. read the letter →

arxiv 2502.05300 v2 pith:MH2EIJU4 submitted 2025-02-07 cs.LG cond-mat.dis-nncs.AIstat.ML

classification cs.LGcond-mat.dis-nncs.AIstat.ML MSC 68T07
keywords parametersymmetrybreakingandrestorationlearningdynamicsmodelcomplexityneuralrepresentationscollapsedeeptheorypositionpaper
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that the fragmented explanations of deep learning can be unified by one mechanism: parameter symmetry breaking and restoration. It claims that three observed hierarchies of neural networks — the abrupt, phase-transition-like jumps in learning dynamics, the adaptive complexity of trained models, and the progressive abstraction of representations — are different manifestations of the same symmetry dynamics. If the position is right, symmetries are not a technical nuisance but a central principle for understanding and intentionally engineering how AI systems learn. The paper is a position piece: it synthesizes existing results, states three concrete hypotheses, and supports each with experiments and theorems drawn from prior work.

What carries the argument

The central object is the parameter symmetry group and the projection matrix $P_G$ onto its invariant subspace. A parameter $\theta$ is in a $G$-symmetric state when $P_G \theta = \theta$; the symmetry-breaking distance $\Delta_G = \|\theta - P_G\theta\|_2^2$ measures how far training has moved the parameters from that state, and the degree of symmetry counts how many pairwise neuron distances exceed a chosen threshold. This gives an operational definition of symmetry breaking and restoration, and it connects to a theorem: at a $G$-symmetric state the effective parameter dimension drops by $\mathrm{rank}(P_G)$, which in the lazy-training regime is equivalent to masking the neural tangent kernel features. The same projection yields a decomposition of the model near a symmetric state into a quadratic kernel model plus a smaller residual model, which is why feature learning is tied to symmetry.

What would settle it

A decisive check is to train the same teacher-student networks with all parameter symmetries removed and ask whether the sharp loss plateaus, complexity jumps, and collapsed representations still appear with unchanged timing; if they do, the claim that symmetry breaking drives the three hierarchies is falsified. A second check is to vary the threshold $\Delta_G^{\mathrm{th}}$ across a wide range and show that the reported coincidence between symmetry-breaking events and loss jumps disappears at alternative reasonable thresholds.

Watch

Extended reading notes

Core claim

The central claim is a unifying hypothesis: the hierarchies of learning dynamics, model complexity, and representation formation are primarily determined by parameter symmetry breaking and restoration. Concretely, the paper argues that the loss landscape of a network is organized into symmetry classes, that gradient-based training moves between these classes, and that each transition changes the effective number of parameters and therefore the model's capacity. It further argues that invariant, hierarchical, and universal representations require parameter symmetry: removing permutation symmetry eliminates neural collapse, while the double rotation symmetry of deep linear networks provably forces universally aligned representations across different models. All three hierarchies are presented as corollaries of the same symmetry mechanism rather than as separate phenomena.

Load-bearing premise

The load-bearing premise is that the hand-chosen threshold $\Delta_G^{\mathrm{th}}$ (0.05–0.2 in most experiments, 1 in one figure) faithfully tracks whether a true group-theoretic symmetry is broken, and that the measured symmetry transitions actually cause the loss jumps and representation changes rather than merely tracking a third factor such as learning-rate or gradient-noise dynamics.

Editorial extensions

If this is right

  • If symmetry dynamics drive learning, then sharp loss drops during training should coincide with symmetry-breaking events, and removing symmetries should eliminate the corresponding plateaus.
  • The space-quantization conjecture implies that under weight decay a layer contains at most a regularizer-dependent number of non-identical neurons, making effective model complexity and generalization essentially width-independent.
  • Neural collapse should disappear when permutation symmetry is removed and reappear when it is restored.
  • The double-rotation symmetry of deep linear networks forces universal alignment of representations across arbitrarily different models, giving a proof of a Platonic-representation-style statement in that setting.
  • Practitioners should be able to engineer desired hierarchies by deliberately introducing or removing symmetries in models, loss functions, and data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If the symmetry-to-symmetry picture transfers from deep linear to nonlinear networks, representation alignment between models may become predictable from the symmetry groups of their architectures, offering a design rule for cross-model compatibility.
  • Editorial inference: The framework suggests a clean experiment: add a single scalar multiplier to a layer's weights (creating one new symmetry) and vary only the regularization strength; if hierarchy-like loss plateaus appear and disappear with that symmetry, the causal story is directly tested.
  • Editorial inference: The threshold-based symmetry metric is a practical proxy; replacing it with a scale-invariant or topologically defined order parameter would make the theory robust to the choice of $\Delta_G^{\mathrm{th}}$.
  • Editorial inference: The space-quantization argument could be extended from neuron weights to attention heads and vocabulary embeddings, predicting that representation rank saturates with width in transformers under weight decay, consistent with the ViT observations in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. This position paper argues that parameter symmetry breaking and restoration is the unifying mechanism behind three empirically observed hierarchies in deep learning: the temporal hierarchy of learning dynamics (e.g., saddle-to-saddle jumps), the complexity hierarchy (e.g., width-independent generalization), and the representational hierarchy (e.g., neural collapse and representation alignment). The paper formalizes parameter symmetries with Definition 1 and Definition 2, defines the symmetry-breaking distance Delta_G in Eq. (3), and presents experimental evidence from MLPs, transformers, ViTs, and ResNets in which symmetry-breaking events are aligned with loss jumps, complexity changes, and representation-rank changes. It also appeals to several theoretical results, including Theorem 1 from the authors' prior work, Theorem 2 from a companion preprint, and Conjecture 1 from an in-preparation manuscript. The stated position is that these three hierarchies are primarily determined by symmetry dynamics, with implications for deliberately engineering symmetries to control learning. The paper itself acknowledges in Section 7 that validating or falsifying its three hypotheses remains the most important next step.

Significance. If the central position is correct, it offers a genuinely unifying research program: distinct phenomena currently explained by separate theories could be understood as different manifestations of a single symmetry-breaking/restoration dynamic, and it would provide a design principle for introducing or removing symmetries in models. The paper is valuable as a synthesis: it connects extensive prior literature on symmetry-induced saddles, stochastic collapse, neural collapse, and representation alignment, and it states three concrete, individually testable hypotheses. Its strengths include explicit operational definitions of symmetry-breaking distance, reproducible experimental protocols in Section A, and honest acknowledgment of alternative views in Section 7. However, the current empirical support is largely correlational and depends on hand-chosen thresholds in the symmetry-breaking measure, and the main theoretical props are informal or unpublished self-references.

major comments (5)
  1. [Section 3, Eq. (3), Figure 2] The operational definition of a symmetry-breaking event is threshold-dependent, and no threshold-stability analysis is reported. The text sets Delta_G_th between 0.05 and 0.2 in most experiments but uses 1.0 in Figure 4, and for the double-rotation symmetry in Figure 3 the measured quantity is the eigenvalue-difference proxy in Eq. (13) rather than Eq. (3). The alignment between symmetry-breaking events and loss jumps or rank changes could therefore be an artifact of threshold choice. The authors should report the detected event times, or the coincidence statistics, over a sweep of Delta_G_th, or use a scale-invariant statistic such as a relative increase in Delta_G. This is load-bearing because all three hypotheses are evaluated with this same measurement style.
  2. [Sections 3-5, central Position statement] The experiments demonstrate temporal coincidence between Delta_G growth and loss drops or representation-rank changes, but coincidence along the same optimization trajectory does not establish that symmetry breaking 'primarily determines' these hierarchies. A common third factor, such as effective learning rate, gradient scale, or stochastic noise, could drive both the symmetry measure and the observed leaps. Figure 5 provides an intervention for the representation hypothesis (removing permutation symmetry suppresses neural collapse), but no analogous intervention is provided for the dynamics and complexity hypotheses. The authors should either soften the 'primarily determined' claim to a correlational one or add intervention experiments, e.g., maintaining the parameters on the symmetric subspace during training or removing/applying individual symmetries and measuring whether the loss jumps and complexity adaptation disappear.
  3. [Section 4 and Appendix B.1, Theorem 1/Theorem 3] The main text claims that at a G-symmetric state the effective model dimension decreases by exactly rank(PG) 'throughout training.' The formal statement in Appendix B.1 shows that the reduced-parameter representation f' exists for all GD/SGD iterations, but the NTK feature-masking statement is proved only in the lazy-training limit (Eqs. (21)-(25) with kappa < 1/(2 lambda_max(A))). The text conflates the exact reduction of parameter count with the stronger kernel-regime claim. This should be flagged explicitly, since Section 4 uses the NTK statement to argue that symmetric states are low-capacity states from which gradient methods cannot escape.
  4. [Section 4, Conjecture 1 and Appendix B.2, Theorem 4] The space-quantization result supporting the Occam's-razor argument relies on the unverified regularity condition in Eq. (26): ||nabla_theta ell_0(...)|| <= K||theta||^q with K = K0 m^{-alpha}. This assumption does the real work in the proof, as Eq. (37) makes clear; without independent evidence that K decays with the number of active neurons for realistic permutable losses, the bound in Eq. (27) is conditional. The authors should prove or at least numerically verify this scaling condition for concrete losses such as MSE with ReLU or attention layers, or otherwise state the result as conditional.
  5. [Sections 4, 5, Appendix B; references [73], [87], [88]] Several load-bearing theoretical items are informal self-references or unpublished: Theorem 1 generalizes the authors' ICLR 2025 paper [88], Theorem 2 is taken from the authors' companion preprint [87], and Conjecture 1 is attributed to an in-preparation paper by two of the authors, cited as 'Removed for anonymity.' For a journal submission, these should be replaced with peer-reviewed versions, precise statements included in the paper, or proofs supplied in the appendix. Reference [73] in particular cannot be verified in its current form and should not be treated as a citable source.
minor comments (5)
  1. [Section 6, Eq. (7)] Equation (7) and the surrounding text contain severe Unicode artifacts (e.g., '⌟⟨⟨⟪rl⟫l⟩⟩⟪...') that make the displayed formula unreadable; this must be fixed before publication.
  2. [Appendix A.1, Eq. (12)-(14)] The notation Ndos is defined as the number of small-Delta_G pairs, but then Ndosb = h - Ndos is called the degree of symmetry breaking; the relationship of these counts to the number of broken generators discussed in Section 3 should be stated explicitly, since h is the number of neurons only for the sorted-neighbor procedure.
  3. [Appendix B.1] Minor typos: 'intialized' should be 'initialized', 'iteraiton' should be 'iteration', and 'kernalized' should be 'kernelized'.
  4. [Figure 2 caption] The caption does not state the threshold Delta_G_th used for the black dotted lines; the reader must infer it from Section A.1. Please report the exact threshold in the caption or in the main text.
  5. [Section 6 and Eq. (4)] The Taylor expansion in Eq. (4) is stated for small Delta_G but the validity condition on Delta_G relative to the Hessian is not discussed; a sentence clarifying the assumed smoothness and the order of the remainder would improve rigor.

Circularity Check

2 steps flagged · score 5.0 of 10

The Complexity and Representation hypotheses rest heavily on the authors' own companion results, including an anonymous in-preparation conjecture, and the empirical 'match' between symmetry-breaking counts and rank/loss is partly a re-thresholding of the same weight geometry; the paper retains independent experiments, so the circularity is real but partial.

  1. self citation load bearing [Section 4 ('Parameter Symmetry as an Occam's Razor'), Conjecture 1; references [73], [87], [88]]
    "Conjecture 1. (Space Quantization Conjecture [73], Informal) In every layer with permutation symmetry, for two nonidentical neurons i and j, ∆ij > O(κβ) after training, where β > 0 and κ is the regularization strength. ... Thus, when regularized, there are, at most, a finite number of nonidentical neurons within a layer, however wide it is. This implies that the actual complexity of the trained network must decrease as a function polynomial in the regularization strength."

    The claim that regularized networks have width-independent complexity is the paper's own Complexity Hypothesis, and it is derived from Conjecture 1, which is attributed to an in-preparation paper by two of the authors ('[73] Yizhou Xu and Liu Ziyin. Removed for anonymity. In Preparation.'). The special-case proof in Section B.2 assumes an unverified scaling K = K0 m^{-alpha} and does not establish the full conjecture. The central complexity conclusion is therefore supported by the authors' own unverified conjecture rather than by an independently checkable external result. The same self-citation pattern appears in the Representation Hypothesis, whose formal support is Theorem 2 imported from the authors' companion preprint [87].

  2. other [Section 5, Figure 4 caption; Appendix A.1 and A.4]
    "This figure repeats the experiment in Ref. [47] and shows the rank and degree of symmetry breaking, which can be seen as the simplest metrics of the representation complexity, of different layers in a 5-layer FCN trained on CIF AR-10. ... As Figure 4 shows, these changes in the representation ranks match well the symmetry-breaking level of the layer, a strong evidence that symmetry may drive the formation of hierarchical representations."

    The paper itself calls both quantities 'simplest metrics of the representation complexity.' Symmetry breaking is operationalized (Section A.1/A.4) as the number of pairwise neuron distances exceeding a hand-chosen threshold, and representation rank is a thresholded count of the same layer's effective dimensions. Because both are monotone summaries of the same weight geometry, their layer-wise correspondence is partly built into the definitions. The alignment is presented as causal evidence that symmetry drives representation hierarchy, but the explanatory variable is a re-thresholding of the quantity being explained, so it cannot independently confirm the Representation Hypothesis.

full rationale

This is a position paper, not a derivation of hard predictions, so much of it is legitimately a synthesis of prior results. The original experiments (Figures 2, 3, 4, 5, 10) give some independent content: for example, removing permutation symmetry changes neural collapse in Figure 5, an external manipulation rather than a definitional identity. However, the formal backbone for the Complexity and Representation hypotheses is unusually dependent on the authors' own prior work: Theorem 1 [88], Theorem 2 [87], Conjecture 1 [73], and the symmetry-removal method [88] all come from the same group, and [73] is even labeled 'Removed for anonymity.' The proof of Theorem 3 in Appendix B.1 and the special case of Conjecture 1 in B.2 are self-contained and reduce the citation dependence, but the full Space Quantization Conjecture remains an unverified self-citation that carries the width-independence claim. Separately, the empirical alignment between 'degree of symmetry breaking' and rank/loss jumps is weakened by the operational definition: thresholded pairwise distances and thresholded rank are two views of the same weight geometry, so a match is not an independent test of causation. These are real circularity concerns, but they are partial; the paper does not reduce entirely to its inputs, and a score of 5 reflects moderate circularity rather than a fully forced derivation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on self-cited companion results, on an empirical threshold for detecting symmetry, and on a regularity assumption used in the quantization proof.

free parameters (2)
  • Delta_G_th (symmetry-breaking distance threshold) = 0.05-0.2 in most experiments; 1.0 in Figure 4
    Hand-chosen cutoff in Definition 2 and Eq. (3) for deciding whether a symmetry is broken; the degree-of-symmetry counts (Ndos) and all experimental conclusions depend on it. No principled derivation is provided.
  • alpha and K0 (scaling exponent and constant in Theorem 4) = unspecified
    Theorem 4 assumes the gradient bound K = K0 m^{-alpha} with free parameters alpha and K0; the resulting bound m <= C gamma^{-(q+1)/(2alpha)} depends on these assumed values, which are not estimated from data.
assumptions (4)
  • domain assumption Parameter symmetries are the primary origins of the saddle points that shape learning dynamics (Refs [41, 83]).
    Underlies the Dynamics Hypothesis in Section 3; not proved in this paper.
  • domain assumption Theorem 2 (informal) from Ref [87] on universal alignment of deep linear networks holds as stated.
    Used to support the Representation Hypothesis in Section 5; no proof is provided in this paper, and the source is a companion preprint by the same authors.
  • domain assumption Inequality (26): the gradient of ell_0 along swap directions is bounded by K times the distance to the q-th power, with K = K0 m^{-alpha}.
    Assumed in Theorem 4 (Appendix B.2) to obtain the space-quantization bound; not verified for realistic losses or architectures.
  • domain assumption The three hierarchy phenomena (complexity jumps, simplicity bias, neural collapse, and related observations) are robust universal behaviors of trained networks.
    The position paper takes these as givens from the cited literature; if some are not robust, the unification claim weakens.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parameter Symmetry Potentially Unifies Deep Learning Theory." pith.science (2026). https://pith.science/paper/MH2EIJU4

@misc{pith2026250205300,
  author       = {Pith},
  title        = {Pith review of: Parameter Symmetry Potentially Unifies Deep Learning Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MH2EIJU4}},
  note         = {Machine review of arXiv:2502.05300}
}
read the original abstract

The dynamics of learning in modern large AI systems is hierarchical, often characterized by abrupt, qualitative shifts akin to phase transitions observed in physical systems. While these phenomena hold promise for uncovering the mechanisms behind neural networks and language models, existing theories remain fragmented, addressing specific cases. In this position paper, we advocate for the crucial role of the research direction of parameter symmetries in unifying these fragmented theories. This position is founded on a centralizing hypothesis for this direction: parameter symmetry breaking and restoration are the unifying mechanisms underlying the hierarchical learning behavior of AI models. We synthesize prior observations and theories to argue that this direction of research could lead to a unified understanding of three distinct hierarchies in neural networks: learning dynamics, model complexity, and representation formation. By connecting these hierarchies, our position paper elevates symmetry -- a cornerstone of theoretical physics -- to become a potential fundamental principle in modern AI.

Figures

Figures reproduced from arXiv: 2502.05300 by the authors.

Figure 1
Figure 1. The division of solution space into hierarchies given by distinct parameter symmetries. Left: Example solution space of a model with parameter symmetries can be divided into hierarchies with boundaries prescribed by symmetry-breaking conditions. The more symmetry there is, the more restricted the hypothesis space becomes. Middle: The temporal (learning dynamics) and spatial (layer-wise information processing) dynami… view at source ↗
Figure 2
Figure 2. DNN learning dynamics is symmetry-to-symmetry. Recent works suggested the learning of neural networks is primarily saddle-to-saddle [33], and escaping these saddle points coincides with a sudden change in the complexity of the network [1]. At the same time, symmetries have been found to be the primary causes of the saddle points [41, 83]. Once the symmetry is removed, saddle points seem to have disappeared when inte… view at source ↗
Figure 3
Figure 3. The complexity and generalization error of neural networks do not grow with width. A well-known observation in deep learning is that overparameterized networks not only work well, but also their generalization errors are empirically found to be essentially independent of width [40, 56, 26, 49], an observation at odds with conventional bounds based on the Rademacher complexity [78]. The existence of parameter symmetr… view at source ↗
Figures from the paper (13 more)
Figure 5
Figure 5. Figure 5: Neural collapse (NC) only happens when permutation symmetry is present. NC is a primary example of how invariant high-level representations emerge in neural networks [54] and exist quite generally in classification, regression, and large language models [5, 71]. When N…
Figure 4
Figure 4. Figure 4: Neural networks learn a hierarchi￾cal representation. Recent works on represen￾tation learning have suggested that the rank of the latent representation first increases and then decreases through the layers [72, 47]. This is reasonable because, on the one hand, a net￾w…
Figure 6
Figure 6. Figure 6: Universally aligned representations emerge in differently trained neural networks. Many recent works demonstrate that different trained neural networks learn representations that are similar to each other and even to the biological brain [30, 36, 75]. Parameter symme￾t…
Figure 7
Figure 7. Figure 7: The same setting as Figure [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The same setting as Figure [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: The same as [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: The degree of symmetry increases with α, the inverse scaling of input (Left), σ, the noise level (Middle), and the degree of teacher symmetry (64−the number of teacher units, Right). Blue dots represent experiment results, and blue lines represent linear fitting. Uppe…
Figure 11
Figure 11. Figure 11: The spectrum of the representation correlation. [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: The same setting as Figure [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: The same setting as Figure [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: The same setting as [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: An extension of [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: The final distance between the weights of neurons for different teachers and the same initialization. [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Observable- and Positional-Encoding-Dependent Symmetry Readout from Neural Network Weights

    cs.LG 2026-07 conditional novelty 6.5 of 10

    Symmetries readable from PE-equipped neural-field weights obey G_obs^exact ⊆ G_lift^exact(φ) ∩ G_true, so PE design structurally gates which groups Gram-type weight observables can detect.

  2. PAC--Bayes Bounds on Quotient Parameter Spaces: Geometry-induced Implicit-Bias Priors

    cs.LG 2026-07 conditional novelty 5.0 of 10

    Identifying parameters that define the same predictor reduces the PAC-Bayes KL complexity term, and a geometry-tilted 'implicit-bias' prior can tighten the certificate when it is closer to the learned posterior.

Reference graph

Works this paper leans on

91 extracted references · 46 canonical work pages · cited by 2 Pith papers

  1. [88]

    Remove symmetries to control model expressivity

    Liu Ziyin, Yizhou Xu, and Isaac Chuang. Remove symmetries to control model expressivity. Interna- tional Conference on Learning Representations , 2025. A Experiments A.1 Measurement of ∆G Permutation Symmetry For the permutation symmetry in fully connected layers, we have described the ∆G for these pairwise symmetries. However, for a layer of width ψ, the...

  2. [87]

    Neural thermodynamics i: Entropic forces in deep and universal representation learning

    Liu Ziyin, Yizhou Xu, and Isaac Chuang. Neural thermodynamics i: Entropic forces in deep and universal representation learning. arXiv preprint arXiv:2505.12387 , 2025

  3. [73]

    Removed for anonymity

    Yizhou Xu and Liu Ziyin. Removed for anonymity. In Preparation

  4. [1]

    Sgd learning on neural networks: leap complexity and saddle-to-saddle dynamics

    Emmanuel Abbe, Enric Boix Adsera, and Theodor Misiakiewicz. Sgd learning on neural networks: leap complexity and saddle-to-saddle dynamics. In The Thirty Sixth Annual Conference on Learning Theory, pages 2552–2623. PMLR, 2023

  5. [2]

    Understanding intermediate layers using linear classifier probes

    Guillaume Alain. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016

  6. [3]

    Negative eigenvalues of the hessian in deep neural networks

    Guillaume Alain, Nicolas Le Roux, and Pierre-Antoine Manzagol. Negative eigenvalues of the hessian in deep neural networks. arXiv preprint arXiv:1902.02366 , 2019

  7. [4]

    More is different: Broken symmetry and the nature of the hierarchical structure of science

    Philip W Anderson. More is different: Broken symmetry and the nature of the hierarchical structure of science. Science, 177(4047):393–396, 1972

  8. [5]

    The prevalence of neural collapse in neural multivariate regression

    George Andriopoulos, Zixuan Dong, Li Guo, Zifan Zhao, and Keith Ross. The prevalence of neural collapse in neural multivariate regression. arXiv preprint arXiv:2409.04180 , 2024

Show all 91 references
  1. [6]

    Thermal forces from a micro- scopic perspective

    Pietro Anzini, Gaia Maria Colombo, Zeno Filiberti, and Alberto Parola. Thermal forces from a micro- scopic perspective. Physical Review Letters, 123(2):028002, 2019. 10

  2. [7]

    Neural networks as kernel learners: The silent alignment effect

    Alexander Atanasov, Blake Bordelon, and Cengiz Pehlevan. Neural networks as kernel learners: The silent alignment effect. arXiv preprint arXiv:2111.00034 , 2021

  3. [8]

    Revisiting model stitching to compare neural representations

    Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. Advances in neural information processing systems , 34:225–236, 2021

  4. [9]

    Nearly-tight vc-dimension and pseudodimension bounds for piecewise linear neural networks

    Peter L Bartlett, Nick Harvey, Christopher Liaw, and Abbas Mehrabian. Nearly-tight vc-dimension and pseudodimension bounds for piecewise linear neural networks. Journal of Machine Learning Research , 20(63):1–17, 2019

  5. [10]

    Representation learning: A review and new perspectives

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence , 35(8):1798–1828, 2013

  6. [11]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇ ckovi´ c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478 , 2021

  7. [12]

    Stochastic collapse: How gradient noise attracts sgd dynamics towards simpler subnetworks

    Feng Chen, Daniel Kunin, Atsushi Yamamura, and Surya Ganguli. Stochastic collapse: How gradient noise attracts sgd dynamics towards simpler subnetworks. arXiv preprint arXiv:2306.04251 , 2023

  8. [13]

    On lazy training in differentiable programming

    Lenaic Chizat, Edouard Oyallon, and Francis Bach. On lazy training in differentiable programming. arXiv preprint arXiv:1812.07956 , 2018

  9. [14]

    Gradient descent on neural networks typically occurs at the edge of stability

    Jeremy M Cohen, Simran Kaur, Yuanzhi Li, J Zico Kolter, and Ameet Talwalkar. Gradient descent on neural networks typically occurs at the edge of stability. arXiv preprint arXiv:2103.00065 , 2021

  10. [15]

    A kernel theory of modern data augmentation

    Tri Dao, Albert Gu, Alexander Ratner, Virginia Smith, Chris De Sa, and Christopher R´ e. A kernel theory of modern data augmentation. InInternational conference on machine learning, pages 1528–1537. PMLR, 2019

  11. [16]

    L. Dinh, R. Pascanu, S. Bengio, and Y. Bengio. Sharp Minima Can Generalize For Deep Nets. ArXiv e-prints, March 2017

  12. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

  13. [18]

    Algorithmic regularization in learning deep homogeneous models: Layers are automatically balanced

    Simon S Du, Wei Hu, and Jason D Lee. Algorithmic regularization in learning deep homogeneous models: Layers are automatically balanced. Advances in neural information processing systems , 31, 2018

  14. [19]

    Why molecules move along a temperature gradient

    Stefan Duhr and Dieter Braun. Why molecules move along a temperature gradient. Proceedings of the National Academy of Sciences , 103(52):19678–19682, 2006

  15. [20]

    The role of permutation invariance in linear mode connectivity of neural networks

    Rahim Entezari, Hanie Sedghi, Olga Saukh, and Behnam Neyshabur. The role of permutation invariance in linear mode connectivity of neural networks. arXiv preprint arXiv:2110.06296 , 2021

  16. [21]

    The representation theory of finite groups

    Walter Feit. The representation theory of finite groups . Elsevier, 1982

  17. [22]

    Spontaneous symmetry breaking in neural networks

    Ricky Fok, Aijun An, and Xiaogang Wang. Spontaneous symmetry breaking in neural networks. arXiv preprint arXiv:1710.06096, 2017

  18. [23]

    A regularity condition of the information matrix of a multilayer perceptron network

    Kenji Fukumizu. A regularity condition of the information matrix of a multilayer perceptron network. Neural networks, 9(5):871–879, 1996

  19. [24]

    Local minima and plateaus in hierarchical structures of multilayer perceptrons

    Kenji Fukumizu and Shun-ichi Amari. Local minima and plateaus in hierarchical structures of multilayer perceptrons. Neural networks, 13(3):317–327, 2000

  20. [25]

    On the role of neural collapse in transfer learning

    Tomer Galanti, Andr´ as Gy¨ orgy, and Marcus Hutter. On the role of neural collapse in transfer learning. arXiv preprint arXiv:2112.15121 , 2021

  21. [26]

    Norm-based generalization bounds for compositionally sparse neural networks

    Tomer Galanti, Mengjia Xu, Liane Galanti, and Tomaso Poggio. Norm-based generalization bounds for compositionally sparse neural networks. arXiv preprint arXiv:2301.12033 , 2023. 11

  22. [27]

    Stochastic training is not necessary for generalization

    Jonas Geiping, Micah Goldblum, Phillip E Pope, Michael Moeller, and Tom Goldstein. Stochastic training is not necessary for generalization. arXiv preprint arXiv:2109.14119 , 2021

  23. [28]

    The implicit bias of depth: How incremental learning drives generalization

    Daniel Gissin, Shai Shalev-Shwartz, and Amit Daniely. The implicit bias of depth: How incremental learning drives generalization. arXiv preprint arXiv:1909.12051 , 2019

  24. [29]

    On the symmetries of deep learn- ing models and their internal representations

    Charles Godfrey, Davis Brown, Tegan Emerson, and Henry Kvinge. On the symmetries of deep learn- ing models and their internal representations. Advances in Neural Information Processing Systems , 35:11893–11905, 2022

  25. [30]

    The platonic representation hypoth- esis

    Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypoth- esis. arXiv preprint arXiv:2405.07987 , 2024

  26. [31]

    Batch normalization: Accelerating deep network training by reduc- ing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reduc- ing internal covariate shift. arXiv preprint arXiv:1502.03167 , 2015

  27. [32]

    Neural tangent kernel: Convergence and general- ization in neural networks

    Arthur Jacot, Franck Gabriel, and Cl´ ement Hongler. Neural tangent kernel: Convergence and general- ization in neural networks. arXiv preprint arXiv:1806.07572 , 2018

  28. [33]

    Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity

    Arthur Jacot, Fran¸ cois Ged, Berfin S ¸im¸ sek, Cl´ ement Hongler, and Franck Gabriel. Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity. arXiv preprint arXiv:2106.15933, 2021

  29. [34]

    Sgd on neural networks learns functions of increasing complexity

    Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. Sgd on neural networks learns functions of increasing complexity. Advances in neural information processing systems , 32, 2019

  30. [35]

    Weight decay induces low-rank attention layers

    Seijin Kobayashi, Yassir Akram, and Johannes Von Oswald. Weight decay induces low-rank attention layers. arXiv preprint arXiv:2410.23819 , 2024

  31. [36]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning , pages 3519–3529. PMLR, 2019

  32. [37]

    Get rich quick: exact solutions reveal how unbalanced initializations promote rapid feature learning

    Daniel Kunin, Allan Ravent´ os, Cl´ ementine Domin´ e, Feng Chen, David Klindt, Andrew Saxe, and Surya Ganguli. Get rich quick: exact solutions reveal how unbalanced initializations promote rapid feature learning. Advances in Neural Information Processing Systems , 37:81157–81...

  33. [38]

    Rethinking the limiting dynamics of sgd: modified loss, phase space oscillations, and anomalous diffusion

    Daniel Kunin, Javier Sagastuy-Brena, Lauren Gillespie, Eshed Margalit, Hidenori Tanaka, Surya Gan- guli, and Daniel LK Yamins. Rethinking the limiting dynamics of sgd: modified loss, phase space oscillations, and anomalous diffusion. 2021

  34. [39]

    Statistical Physics: Volume 5 , volume 5

    Lev Davidovich Landau and Evgenii Mikhailovich Lifshitz. Statistical Physics: Volume 5 , volume 5. Elsevier, 2013

  35. [40]

    Neural network renormalization group

    Shuo-Hui Li and Lei Wang. Neural network renormalization group. Physical review letters , 121(26):260601, 2018

  36. [41]

    Symmetry, saddle points, and global geometry of nonconvex matrix factorization

    Xingguo Li, Zhaoran Wang, Junwei Lu, Raman Arora, Jarvis Haupt, Han Liu, and Tuo Zhao. Symmetry, saddle points, and global geometry of nonconvex matrix factorization. arXiv preprint arXiv:1612.09296, 1:5–1, 2016

  37. [42]

    Reconciling modern deep learning with traditional op- timization analyses: The intrinsic learning rate

    Zhiyuan Li, Kaifeng Lyu, and Sanjeev Arora. Reconciling modern deep learning with traditional op- timization analyses: The intrinsic learning rate. Advances in Neural Information Processing Systems , 33:14544–14555, 2020

  38. [43]

    What happens after sgd reaches zero loss?–a mathe- matical framework

    Zhiyuan Li, Tianhao Wang, and Sanjeev Arora. What happens after sgd reaches zero loss?–a mathe- matical framework. In International Conference on Learning Representations , 2021

  39. [44]

    The empirical impact of neural parameter symmetries, or lack thereof

    Derek Lim, Theo Moe Putterman, Robin Walters, Haggai Maron, and Stefanie Jegelka. The empirical impact of neural parameter symmetries, or lack thereof. arXiv preprint arXiv:2405.20231 , 2024. 12

  40. [45]

    Abide by the law and follow the flow: Conserva- tion laws for gradient flows

    Sibylle Marcotte, R´ emi Gribonval, and Gabriel Peyr´ e. Abide by the law and follow the flow: Conserva- tion laws for gradient flows. 2023

  41. [46]

    Invariant and equivariant graph networks

    Haggai Maron, Heli Ben-Hamu, Nadav Shamir, and Yaron Lipman. Invariant and equivariant graph networks. arXiv preprint arXiv:1812.09902 , 2018

  42. [47]

    The tunnel effect: Building data representations in deep neural networks

    Wojciech Masarczyk, Mateusz Ostaszewski, Ehsan Imani, Razvan Pascanu, Piotr Mi lo´ s, and Tomasz Trzcinski. The tunnel effect: Building data representations in deep neural networks. Advances in Neural Information Processing Systems, 36, 2024

  43. [48]

    Charge symmetry, quarks and mesons

    Gerald A Miller, Bernard Marie Karel Nefkens, and Ivo ˇSlaus. Charge symmetry, quarks and mesons. Physics Reports, 194(1-2):1–116, 1990

  44. [49]

    Deep neural networks have an inbuilt occam’s razor

    Chris Mingard, Henry Rees, Guillermo Valle-P´ erez, and Ard A Louis. Deep neural networks have an inbuilt occam’s razor. Nature Communications, 16(1):220, 2025

  45. [50]

    Towards understanding the role of over-parametrization in generalization of neural networks

    Behnam Neyshabur, Zhiyuan Li, Srinadh Bhojanapalli, Yann LeCun, and Nathan Srebro. Towards understanding the role of over-parametrization in generalization of neural networks. arXiv preprint arXiv:1805.12076, 2018

  46. [51]

    In search of the real inductive bias: On the role of implicit regularization in deep learning

    Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614 , 2014

  47. [52]

    On connected sublevel sets in deep learning

    Quynh Nguyen. On connected sublevel sets in deep learning. In International conference on machine learning, pages 4790–4799. PMLR, 2019

  48. [53]

    Neural networks should be wide enough to learn disconnected decision regions

    Quynh Nguyen, Mahesh Chandra Mukkamala, and Matthias Hein. Neural networks should be wide enough to learn disconnected decision regions. In International conference on machine learning , pages 3740–3749. PMLR, 2018

  49. [54]

    Prevalence of neural collapse during the terminal phase of deep learning training

    Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020

  50. [55]

    An introduction to quantum field theory

    Michael E Peskin. An introduction to quantum field theory . CRC press, 2018

  51. [56]

    On generalization bounds for neural networks with low rank layers

    Andrea Pinto, Akshay Rangamani, and Tomaso Poggio. On generalization bounds for neural networks with low rank layers. arXiv preprint arXiv:2411.13733 , 2024

  52. [57]

    Cosmology and broken discrete symmetry

    John Preskill, Sandip P Trivedi, Frank Wilczek, and Mark B Wise. Cosmology and broken discrete symmetry. Nuclear Physics B , 363(1):207–220, 1991

  53. [58]

    Neural collapse in deep homogeneous classifiers and the role of weight decay

    Akshay Rangamani and Andrzej Banburski-Fahey. Neural collapse in deep homogeneous classifiers and the role of weight decay. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 4243–4247. IEEE, 2022

  54. [59]

    Feature learning in deep classifiers through intermediate neural collapse

    Akshay Rangamani, Marius Lindegaard, Tomer Galanti, and Tomaso A Poggio. Feature learning in deep classifiers through intermediate neural collapse. In International Conference on Machine Learning, pages 28729–28745. PMLR, 2023

  55. [60]

    Exact solutions to the nonlinear dynamics of learning in deep linear neural networks

    Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120 , 2013

  56. [61]

    On the stepwise nature of self-supervised learning

    James B Simon, Maksis Knutins, Liu Ziyin, Daniel Geisz, Abraham J Fetterman, and Joshua Albrecht. On the stepwise nature of self-supervised learning. arXiv preprint arXiv:2303.15438 , 2023

  57. [62]

    Geometry of the loss landscape in overparameterized neural networks: Symmetries and invariances

    Berfin Simsek, Fran¸ cois Ged, Arthur Jacot, Francesco Spadaro, Cl´ ement Hongler, Wulfram Gerstner, and Johanni Brea. Geometry of the loss landscape in overparameterized neural networks: Symmetries and invariances. In International Conference on Machine Learning , pages 9722–...

  58. [63]

    On the origin of implicit regular- ization in stochastic gradient descent

    Samuel L Smith, Benoit Dherin, David GT Barrett, and Soham De. On the origin of implicit regular- ization in stochastic gradient descent. arXiv preprint arXiv:2101.12176 , 2021. 13

  59. [64]

    Noether’s learning dynamics: Role of symmetry breaking in neural networks

    Hidenori Tanaka and Daniel Kunin. Noether’s learning dynamics: Role of symmetry breaking in neural networks. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 25646–25660....

  60. [65]

    Optimizing mode connectivity via neuron alignment.Advances in Neural Information Processing Systems, 33:15300– 15311, 2020

    Norman Tatro, Pin-Yu Chen, Payel Das, Igor Melnyk, Prasanna Sattigeri, and Rongjie Lai. Optimizing mode connectivity via neuron alignment.Advances in Neural Information Processing Systems, 33:15300– 15311, 2020

  61. [66]

    Equivalences between sparse models and neural networks

    Ryan J Tibshirani. Equivalences between sparse models and neural networks. Working Notes. URL https://www. stat. cmu. edu/˜ ryantibs/papers/sparsitynn. pdf , 2021

  62. [67]

    The information bottleneck method

    Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000

  63. [68]

    Deep learning and the information bottleneck principle

    Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle. In 2015 ieee information theory workshop (itw) , pages 1–5. IEEE, 2015

  64. [69]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

  65. [70]

    Asymptotic equivalence of bayes cross validation and widely applicable information criterion in singular learning theory

    Sumio Watanabe and Manfred Opper. Asymptotic equivalence of bayes cross validation and widely applicable information criterion in singular learning theory. Journal of machine learning research , 11(12), 2010

  66. [71]

    Linguistic collapse: Neural collapse in (large) language models

    Robert Wu and Vardan Papyan. Linguistic collapse: Neural collapse in (large) language models. arXiv preprint arXiv:2405.17767, 2024

  67. [72]

    The janus effects of sgd vs gd: high noise and low rank

    Mengjia Xu, Tomer Galanti, Akshay Rangamani, Lorenzo Rosasco, and Tomaso Poggio. The janus effects of sgd vs gd: high noise and low rank. 2023

  68. [74]

    When does feature learning happen? perspective from an analytically solvable model

    Yizhou Xu and Liu Ziyin. When does feature learning happen? perspective from an analytically solvable model. arXiv preprint arXiv:2401.07085 , 2024

  69. [75]

    Performance-optimized hierarchical models predict neural responses in higher visual cortex

    Daniel LK Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the national academy of sciences , 111(23):8619–8624, 2014

  70. [76]

    Feature learning in infinite-width neural networks

    Greg Yang and Edward J Hu. Feature learning in infinite-width neural networks. arXiv preprint arXiv:2011.14522, 2020

  71. [77]

    Visualizing and understanding convolutional networks

    Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceed- ings, Part I 13 , pages 818–833. Springer, 2014

  72. [78]

    Understanding deep learning requires rethinking generalization

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. 2017

  73. [79]

    Symmetry teleportation for accelerated opti- mization

    Bo Zhao, Nima Dehmamy, Robin Walters, and Rose Yu. Symmetry teleportation for accelerated opti- mization. Advances in Neural Information Processing Systems , 35:16679–16690, 2022

  74. [80]

    Understanding mode connectivity via param- eter space symmetry

    Bo Zhao, Nima Dehmamy, Robin Walters, and Rose Yu. Understanding mode connectivity via param- eter space symmetry. In UniReps: the First Workshop on Unifying Representations in Neural Models , 2023

  75. [81]

    Improving convergence and generalization using parameter symmetries

    Bo Zhao, Robert M Gower, Robin Walters, and Rose Yu. Improving convergence and generalization using parameter symmetries. arXiv preprint arXiv:2305.13404 , 2023. 14

  76. [82]

    Quadratic models for understanding neural network dynamics

    Libin Zhu, Chaoyue Liu, Adityanarayanan Radhakrishnan, and Mikhail Belkin. Quadratic models for understanding neural network dynamics. arXiv preprint arXiv:2205.11787 , 2022

  77. [83]

    Symmetry induces structure and constraint of learning

    Liu Ziyin. Symmetry induces structure and constraint of learning. In Forty-first International Confer- ence on Machine Learning , 2024

  78. [84]

    Formation of representations in neural networks

    Liu Ziyin, Isaac Chuang, Tomer Galanti, and Tomaso Poggio. Formation of representations in neural networks. International Conference on Learning Representations , 2025

  79. [85]

    The probabilistic stability of stochastic gradient descent, 2023

    Liu Ziyin, Botao Li, Tomer Galanti, and Masahito Ueda. The probabilistic stability of stochastic gradient descent, 2023

  80. [86]

    Parameter symmetry and noise equilibrium of stochastic gradient descent

    Liu Ziyin, Mingze Wang, Hongchao Li, and Lei Wu. Parameter symmetry and noise equilibrium of stochastic gradient descent. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  81. [89]

    For all time steps t under GD or SGD, there exists a model f′(x, θ′) and sequence of parameters θ′ t such that for all x, f′(x, θ′ t)= f(x, θt), (18) where dim(θ′)= dim(PG)

  82. [90]

    Here, A ∶= PG ∑x∇θf(x, θ0)T∇θf(x, θ0)PG and A+ denotes the Moore–Penrose inverse of A

    The kernalized model, g(x, θ)= limλ→0(λ−1f(x, λθ+ θ0)− f(x, θ0)), converges to θ∗ = A+ ∑ x ∇θf(x, θ0)T y(x) (19) under GD for a sufficiently small learning rate. Here, A ∶= PG ∑x∇θf(x, θ0)T∇θf(x, θ0)PG and A+ denotes the Moore–Penrose inverse of A. The second part of the theor...

  83. [91]

    (21) Therefore, g(x, θ) simplifies to a kernel model g(x, θ)=∇θ0 f(x, θ0)PGθ

    By (20), close to any symmetric point θ0 (any θ0 for which PGθ0 = θ0), for all x, we have f(x, θ)− f(x, θ0)=∇θf(x, θ0)PG∆+ O(∥∆∥)2. (21) Therefore, g(x, θ) simplifies to a kernel model g(x, θ)=∇θ0 f(x, θ0)PGθ. (22) Let us consider the squared lossℓ(θ)= ∑x∣∣y(x)−g(x, θ)∣∣2 and ...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.